arXiv subjects

Lawrence Hall

Publications and source records attributed to Lawrence Hall.

At least 19 recordsLinked to original sources

arXiv2025

Noise Injection: Improving Out-of-Distribution Generalization for Limited Size Datasets

Deep learned (DL) models for image recognition have been shown to fail to generalize to data from different devices, populations, etc. COVID-19 detection from Chest X-rays (CXRs), in particular, has been shown to fail to generalize to out-of-distribution (OOD) data from new clinical sources not covered in the training set. This occurs because models learn to exploit shortcuts - source-specific artifacts that do not translate to new distributions - rather than reasonable biomarkers to maximize performance on in-distribution (ID) data. Rendering the models more robust to distribution shifts, our study investigates the use of fundamental noise injection techniques (Gaussian, Speckle, Poisson, and Salt and Pepper) during training. Our empirical results demonstrate that this technique can significantly reduce the performance gap between ID and OOD evaluation from 0.10-0.20 to 0.01-0.06, based on results averaged over ten random seeds across key metrics such as AUC, F1, accuracy, recall and specificity. Our source code is publicly available at https://github.com/Duongmai127/Noisy-ood

cs.CV↗

arXiv2025

Integral Signatures of Activation Functions: A 9-Dimensional Taxonomy and Stability Theory for Deep Learning

Activation functions govern the expressivity and stability of neural networks, yet existing comparisons remain largely heuristic. We propose a rigorous framework for their classification via a nine-dimensional integral signature S_sigma(phi), combining Gaussian propagation statistics (m1, g1, g2, m2, eta), asymptotic slopes (alpha_plus, alpha_minus), and regularity measures (TV(phi'), C(phi)). This taxonomy establishes well-posedness, affine reparameterization laws with bias, and closure under bounded slope variation. Dynamical analysis yields Lyapunov theorems with explicit descent constants and identifies variance stability regions through (m2', g2). From a kernel perspective, we derive dimension-free Hessian bounds and connect smoothness to bounded variation of phi'. Applying the framework, we classify eight standard activations (ReLU, leaky-ReLU, tanh, sigmoid, Swish, GELU, Mish, TeLU), proving sharp distinctions between saturating, linear-growth, and smooth families. Numerical Gauss-Hermite and Monte Carlo validation confirms theoretical predictions. Our framework provides principled design guidance, moving activation choice from trial-and-error to provable stability and kernel conditioning.

cs.LG↗

arXiv2024

Strong CP and Flavor in Multi-Higgs Theories

We introduce a class of multi-Higgs doublet extensions of the Standard Model that solves the strong CP problem with profound consequences for the flavor sector. The Yukawa matrices are constrained to have many zero entries by a "Higgs-Flavor" symmetry, $G_{\rm HF}$, that acts on Higgs and quark fields. The violation of both CP and $G_{\rm HF}$ occurs in the Higgs mass matrix so that, for certain choices of $G_{\rm HF}$ charges, the strong CP parameter $\bar{\theta}$ is zero at tree-level. Radiative corrections to $\bar{\theta}$ are computed in this class of theories. They vanish in realistic two-Higgs doublet models with $G_{\rm HF} = \mathbb{Z}_3$. We also construct realistic three-Higgs models with $G_{\rm HF} = \rm U(1)$, where the one-loop results for $\bar{\theta}$ are model-dependent. Requiring $\bar{\theta}< 10^{-10}$ has important implications for the flavor problem by constraining the Yukawa coupling and Higgs mass matrices. Contributions to $\bar{\theta}$ from higher-dimension operators are computed at 1-loop and can also be sufficiently small, although the hierarchy problem of this class of theories is worse than in the Standard Model.

hep-ph↗

arXiv2024

A Preliminary Study on Using Large Language Models in Software Pentesting

Large language models (LLM) are perceived to offer promising potentials for automating security tasks, such as those found in security operation centers (SOCs). As a first step towards evaluating this perceived potential, we investigate the use of LLMs in software pentesting, where the main task is to automatically identify software security vulnerabilities in source code. We hypothesize that an LLM-based AI agent can be improved over time for a specific security task as human operators interact with it. Such improvement can be made, as a first step, by engineering prompts fed to the LLM based on the responses produced, to include relevant contexts and structures so that the model provides more accurate results. Such engineering efforts become sustainable if the prompts that are engineered to produce better results on current tasks, also produce better results on future unknown tasks. To examine this hypothesis, we utilize the OWASP Benchmark Project 1.2 which contains 2,740 hand-crafted source code test cases containing various types of vulnerabilities. We divide the test cases into training and testing data, where we engineer the prompts based on the training data (only), and evaluate the final system on the testing data. We compare the AI agent's performance on the testing data against the performance of the agent without the prompt engineering. We also compare the AI agent's results against those from SonarQube, a widely used static code analyzer for security testing. We built and tested multiple versions of the AI agent using different off-the-shelf LLMs -- Google's Gemini-pro, as well as OpenAI's GPT-3.5-Turbo and GPT-4-Turbo (with both chat completion and assistant APIs). The results show that using LLMs is a viable approach to build an AI agent for software pentesting that can improve through repeated use and prompt engineering.

cs.CR↗

arXiv2023

Dark Matter in A Mirror Solution to the Strong CP Problem

We study thermal production of dark matter (DM) in a realization of the minimal models of Ref.~\cite{Bonnefoy:2023afx}, where parity is used to solve the strong CP problem by transforming the entire Standard Model (SM) into a mirror copy. Although the mirror electron $e^{\prime}$ is a good DM candidate, its viability is mired by the presence of the mirror up-quark $u^{\prime}$, whose abundance is intimately related to the $e^{\prime}$ abundance and must be suppressed. This can be achieved through a sequential freeze-in mechanism, where mirror photons are first produced from SM gluons, and then the mirror photons produce $e'$. After computing the details of this double freeze-in, we discuss the allowed parameter space of the model, which lies at the threshold of experimental observations. We find that this origin of $e'$ DM requires a low reheating temperature after inflation and is consistent with the baryon asymmetry arising from leptogenesis, providing mirror neutrinos have a significant degeneracy. Finally, we show that this $e'$ DM is not compatible with Higgs Parity, the simplest scheme with exact parity, unless SM parameters deviate significantly from their central values or the minimal model is extended.

hep-ph↗

arXiv2023

A Colorful Mirror Solution to the Strong CP Problem

We propose theories of a complete mirror world with parity (P) solving the strong CP problem. P exchanges the entire Standard Model (SM) with its mirror copy. We derive bounds on the two new mass scales that arise: $v'$ where parity and mirror electroweak symmetry are spontaneously broken, and $v_3$ where the color groups break to the diagonal strong interactions. The strong CP problem is solved even if $v_3 \ll v^{\prime}$, when heavy coloured states at the scale $v_3$ may be accessible at LHC and future colliders. Furthermore, we argue that the breaking of P introduces negligible contributions to $\bar \theta_\text{QCD}$, starting at three-loop order. The symmetry breaking at $v_3$ can be made dynamical, without introducing an additional hierarchy problem.

hep-ph↗

arXiv2022

Simulating User-Level Twitter Activity with XGBoost and Probabilistic Hybrid Models

The Volume-Audience-Match simulator, or VAM was applied to predict future activity on Twitter related to international economic affairs. VAM was applied to do timeseries forecasting to predict the: (1) number of total activities, (2) number of active old users, and (3) number of newly active users over the span of 24 hours from the start time of prediction. VAM then used these volume predictions to perform user link predictions. A user-user edge was assigned to each of the activities in the 24 future timesteps. VAM considerably outperformed a set of baseline models in both the time series and user-assignment tasks

cs.LG↗

arXiv2013

Why Comparable? A Multiverse Explanation of the Dark Matter-Baryon Coincidence

The densities of dark and baryonic matter are comparable: ζ= ρ_D / ρ_B ~ O(1). This is surprising because they are controlled by different combinations of low-energy physics parameters. Here we consider the probability distribution over ζin the landscape. We argue that the Why Comparable problem can be solved without detailed anthropic assumptions, and independently of the nature of dark matter. Overproduction of dark matter suppresses the probability like 1/(1+ζ), if the causal patch is used to regulate infinities. This suppression can counteract a prior distribution favoring large ζ, selecting ζ~ O(1). This effect not only explains the Why Comparable coincidence but also renders otherwise implausible models of dark matter viable. For the special case of axion dark matter, Wilczek and independently Freivogel have already noted that a 1/(1+ζ) suppression prevents overproduction of a GUT-scale QCD axion. If the dark matter is the LSP, the effect can explain the moderate fine-tuning of the weak scale in simple supersymmetric models.

hep-th↗

arXiv2011

Gravitino Freeze-In

We explore an alternative mechanism for the production of gravitino dark matter whereby relic gravitinos originate from the decays of superpartners which are still in thermal equilibrium, i.e. via freeze-in. Contributions to the gravitino abundance from freeze-in can easily dominate over those from thermal scattering over a broad range of parameter space, e.g. when the scalar superpartners are heavy. Because the relic abundance from freeze-in is independent of the reheating temperature after inflation, collider measurements may be used to unambiguously reconstruct the freeze-in origin of gravitinos. In particular, if gravitino freeze-in indeed accounts for the present day dark matter abundance, then the lifetime of the next-to-lightest superpartner is uniquely fixed by the superpartner spectrum.

hep-ph↗

arXiv2001

SO(10) Unified Theories in Six Dimensions

We construct supersymmetric models of SO(10) unification in which the gauge symmetry is broken by orbifold compactification. We find that using boundary conditions to break the gauge symmetry down to $SU(3)_C \otimes SU(2)_L \otimes U(1)_Y \otimes U(1)_X$ without leaving unwanted massless states requires at least two extra dimensions, motivating us to work with 6D orbifolds. SO(10) is broken by two operations, each of which induces gauge-breaking to either the Georgi-Glashow, Pati-Salam, or flipped $SU(5) \otimes U(1)$ subgroups; assigning different unbroken subgroups to the two operations leaves only the standard model gauge group and $U(1)_X$ unbroken. The models we build employ extra-dimensional mechanisms for naturally realizing doublet-triplet splitting, suppressing proton decay, and avoiding unwanted grand-unified fermion mass relations. We find some tension between being free of anomalies of the 6D bulk, accommodating a simple mechanism for generating right-handed neutrino masses, and preserving the precise prediction of the weak mixing angle.

hep-ph↗

arXiv2001

Gauge-Higgs Unification in Higher Dimensions

The electroweak Higgs doublets are identified as components of a vector multiplet in a higher dimensional supersymmetric field theory. We construct a minimal model in 6D where the electroweak $SU(2) \otimes U(1)$ gauge group is extended to SU(3), and unified 6D models with the unified SU(5) gauge symmetry extended to SU(6). In these realistic theories the extended gauge group is broken by orbifold boundary conditions, leaving Higgs doublet zero modes which have Yukawa couplings to quarks and leptons on the orbifold fixed points. In one SU(6) model the weak mixing angle receives power law corrections, while in another the fixed point structure forbids such corrections. A 5D model is also constructed in which the Higgs doublet contains the fifth component of the gauge field. In this case Yukawa couplings are introduced as non-local operators involving the Wilson line of this gauge field.

hep-ph↗

arXiv2001

Towards a Theory of Flavor from Orbifold GUTs

We show that the recently constructed 5-dimensional supersymmetric $S^1/(Z_2\times Z_2')$ orbifold GUT models allow an appealing explanation of the observed hierarchical structure of the quark and lepton masses and mixing angles. Flavor hierarchies arise from the geometrical suppression of some couplings when fields propagate in different numbers of dimensions, or on different fixed branes. Restrictions arising from locality in the extra dimension allow interesting texture zeroes to be easily generated. In addition the detailed nature of the SU(5)-breaking orbifold projections lead to simple theories where $b-\tau$ unification is maintained but similar disfavored SU(5) relations for the lighter generations are naturally avoided. We find that simple 5d models based on $S^1/(Z_2\times Z_2')$ are strikingly successful in explaining many features of the masses and mixing angles of the 2nd and 3rd generation. Successful three generation models of flavor including neutrinos are constructed by generalizing the $S^1/(Z_2\times Z'_2)$ model to six dimensions. Large angle neutrino mixing is elegantly accommodated. Novel features of these models include a simple $m_u=0$ configuration leading to a solution of the strong CP problem.

hep-ph↗

arXiv2001

Gauge Unification in Higher Dimensions

A complete 5-dimensional SU(5) unified theory is constructed which, on compactification on the orbifold with two different Z_2's (Z_2 and Z_2'), yields the minimal supersymmetric standard model. The orbifold accomplishes SU(5) gauge symmetry breaking, doublet-triplet splitting, and a vanishing of proton decay from operators of dimension 5. Until 4d supersymmetry is broken, all proton decay from dimension 4 and dimension 5 operators is forced to vanish by an exact U(1)_R symmetry. Quarks and leptons and their Yukawa interactions are located at the Z_2 orbifold fixed points, where SU(5) is unbroken. A new mechanism for introducing SU(5) breaking into the quark and lepton masses is introduced, which originates from the SU(5) violation in the zero-mode structure of bulk multiplets. Even though SU(5) is absent at the Z_2' orbifold fixed point, the brane threshold corrections to gauge coupling unification are argued to be negligibly small, while the logarithmic corrections are small and in a direction which improves the agreement with the experimental measurements of the gauge couplings. Furthermore, the X gauge boson mass is lowered, so that proton decay to e^+ π^0 is expected with a rate within about one order of magnitude of the current limit. Supersymmetry breaking occurs on the Z_2' orbifold fixed point, and is felt directly by the gauge and Higgs sectors, while squarks and sleptons acquire mass via gaugino mediation, solving the supersymmetric flavor problem.

hep-ph↗

arXiv2001

Finite Radiative Electroweak Symmetry Breaking from the Bulk

A new physical origin for electroweak symmetry breaking is proposed, involving compact spatial dimensions of scale 1/R \approx 1 TeV. The higher dimensional theory is supersymmetric, and hence requires the top-quark Yukawa coupling to be localized on some ``Yukawa brane'' in the bulk. The short distance divergence in the Higgs-boson mass is regulated because supersymmetry is unbroken in the vicinity of this Yukawa brane. A finite, negative Higgs mass-squared is generated radiatively by the top-quark supermultiplet propagating a distance of order R from the Yukawa brane to probe supersymmetry breaking. The physics of electroweak symmetry breaking is therefore closely related to this top propagation across the bulk, and is dominated by the mass scale 1/R, with exponential insensitivity to higher energy scales. The masses of the superpartners and the Kaluza-Klein resonances are also set by the mass scale 1/R, which is naturally larger than the W boson mass by a loop factor. Explicit models are constructed which are highly constrained and predictive. The finite radiative correction to the Higgs mass is computed, and the Higgs sector briefly explored. The superpartner and Kaluza-Klein resonance spectra are calculated, and the problem of flavor violation from squark and slepton exchange is solved. Important collider signatures include highly ionizing charged tracks from stable top squarks, and events with two Higgs bosons and missing transverse energy.

hep-ph↗

arXiv2000

Small Neutrino Masses from Supersymmetry Breaking

An alternative to the conventional see-saw mechanism is proposed to explain the origin of small neutrino masses in supersymmetric theories. The masses and couplings of the right-handed neutrino field are suppressed by supersymmetry breaking, in a way similar to the suppression of the Higgs doublet mass, $μ$. New mechanisms for light Majorana, Dirac and sterile neutrinos arise, depending on the degree of suppression. Superpartner phenomenology is greatly altered by the presence of weak scale right-handed sneutrinos, which may have a coupling to a Higgs boson and a left-handed sneutrino. The sneutrino spectrum and couplings are quite unlike the conventional case - the lightest sneutrino can be the dark matter and predictions are given for event rates at upcoming halo dark matter direct detection experiments. Higgs decays and search strategies are changed. Copious Higgs production at hadron colliders can result from cascade decays of squarks and gluinos.

hep-ph↗

arXiv2000

Neutrino Masses at v^{3/2}

Theories in which neutrino masses are generated by a conventional see-saw mechanism generically yield masses which are O(v^2) in units where M_{Pl}=1, which is naively too small to explain the results from SuperKamiokande. In supersymmetric theories with gravity mediated supersymmetry breaking, the fundamental small parameter is not v/M_{Pl}, but m_I/M_{Pl}, where m_I is the scale of supersymmetry breaking in the hidden sector. We note that m_I^3/M_{Pl}^2 is only slightly too large to explain SuperKamiokande, and present two models that achieve neutrino masses at this order in m_I, one of which has an additional suppression lambda_tau^2, while the other has additional suppression arising from a loop factor. The latter model shares a great deal of phenomenology with a class of models previously explored, including the possibility of viable sneutrino dark matter.

hep-ph↗

arXiv2000

Solving the Hierarchy Problem with Exponentially Large Dimensions

In theories with (sets of) two large extra dimensions and supersymmetry in the bulk, the presence of non-supersymmetric brane defects naturally induces a logarithmic potential for the volume of the transverse dimensions. Since the logarithm of the volume rather than the volume itself is the natural variable, parameters of O(10) in the potential can generate an exponentially large size for the extra dimensions. This provides a true solution to the hierarchy problem, on the same footing as technicolor or dynamical supersymmetry breaking. The area moduli have a Compton wavelength of about a millimeter and mediate Yukawa interactions with gravitational strength. We present a simple explicit example of this idea which generates two exponentially large dimensions. In this model, the area modulus mass is in the millimeter range even for six dimensional Planck scales as high as 100 TeV.

hep-ph↗

arXiv1999

Exponentially Small Supersymmetry Breaking from Extra Dimensions

The supersymmetric ``shining'' of free massive chiral superfields in extra dimensions from a distant source brane can trigger exponentially small supersymmetry breaking on our brane of order e^{-2 pi R}, where R is the radius of the extra dimensions. This supersymmetry breaking can be transmitted to the superpartners in a number of ways, for instance by gravity or via the standard model gauge interactions. The radius R can easily be stabilized at a size O(10) larger that the fundamental scale. The models are extremely simple, relying only on free, classical bulk dynamics to solve the hierarchy problem.

hep-ph↗